Machine learning surrogates for stability assessment of dry-stacked concrete waste assemblies: Comparing ensemble and graph-based architectures

The integration of non-standard materials into construction workflows is a key step toward circularity in the built environment. However, the lack of rapid stability assessment methods for non-standard assemblies limits their adoption, particularly in early-stage design when the potential for reducing environmental impact is greatest. This paper presents a comparative analysis of physics-informed surrogate model architectures and feature engineering strategies to sidestep the computational latency of physics simulations for dry-stacked concrete waste assemblies. We benchmark ensemble tree-based methods (RF, XGBoost) against four Graph Neural Network (GNN) variants across three feature configurations: physics-derived features, masonry rule features, and their combination. A dataset of 10,654 dry-stacked concrete wall assemblies is generated through probabilistic waste inventories using 22 packing heuristics, with stability labels derived from rigid body simulation. The models are trained to predict average displacement under gravitational loading, bypassing costly simulations at inference time. Results demonstrate that predictive accuracy is contingent upon the alignment between model inductive bias and feature representation rather than model complexity. While GNNs significantly outperform tree-based methods on single-category features by leveraging contact graph topology, both families converge on the combined feature set (R 2 =0.714). Notably, an uncertainty-weighted GNN ensemble achieves peak accuracy (R 2 =0.740), while packing heuristic selection emerges as a first-order design variable influencing stability by over 100%. This framework establishes that domain-aware feature engineering can equalize performance across architectures, enabling millisecond-scale feedback for real-time, performance-driven exploration of reclaimed material inventories.